Open-Weight Models

DPO checkpoints from a study of why DPO makes Olmo 3 7B Instruct overconfident on multiple-choice questions. Every model here is trained from allenai/Olmo-3-7B-Instruct-SFT on Ai2's Dolci-Instruct-DPO preference pairs with Ai2's open-instruct code. - Code, results and the write-up ("Why Does DPO Make Olmo 3 Overconfident?"): These are research checkpoints for reproducing the study's measurements. They were not tuned or safety-tested for use. "Agreement-weighted" DPO multiplies each pair's loss by sigmoid(-5 x the SFT model's preference margin), rescaled to average 1, so pairs the SFT model already gets right count less. Each folder is a full model (model.safetensors, config, tokenizer, chat…

static quants of https://huggingface.co/gh0st359/omida-r1

Code, logs, results and training checkpoints of OmniJev / MM-Jev: Jev-style typed decisions (noul / choice / score) on Gemma 3n E4B with text, image, audio and video states. This is a scratch workspace synced by the training job; the finished model is published separately. code/ – training / evaluation code (MM-Jev architecture: option tree in one pass, KV-share truncation, modality exit + latent memory, sibling bridge, MatFormer width slicing) run/ – pipeline log, results.json, checkpoints (LoRA + decision heads + latents) Base model: google/gemma-3n-E4B-it (Gemma terms of use apply to derivatives).

Re-exported llmdecoder sub-model for OmniVoice Web, a fully in-browser (WebGPU) TypeScript runtime for k2-fsa/OmniVoice. The public onnx-community/OmniVoice-Onnx llmdecoder uses causal attention (GroupQueryAttention + 2-D causal mask). OmniVoice is a bidirectional masked-diffusion LM, so with causal attention the model ignores the text prompt and emits generic speech. This repo hosts the Qwen3-0.6B core re-exported with a 4-D all-zero (non-causal) attention mask, which restores exact text conditioning. All three reproduce the PyTorch reference with WER 0 (round-trip Whisper) on the evaluation sentence. The embeddings/heads/tokenizer/codec sub-models remain on tools/exportllm.py --dtype…

ACT policy trained on 200 demonstration episodes. Selected checkpoint: 150,000 training steps. Matching hardware calibration and custom LeRobot code are required. Keep all processor JSON and normalization safetensors files together with the model weights. Source run: warehouseall200twocyclesv3act250k. Selected through operator tests; no standardized success rate is claimed. Deployment code: https://github.com/E1I6-Logistics/OMX (access may be required). Base PC source: LeRobot 0.6.2, commit 89236ea0f4f81a81ca566081e20dd1ff5f823cbe plus custom modifications.

ACT policy trained on 100 demonstration episodes. Selected checkpoint: 150,000 training steps. Matching hardware calibration and custom LeRobot code are required. Keep all processor JSON and normalization safetensors files together with the model weights. Source run: warehouseitemb100mergedact200k. Selected through operator tests; no standardized success rate is claimed. Deployment code: https://github.com/E1I6-Logistics/OMX (access may be required). Base PC source: LeRobot 0.6.2, commit 89236ea0f4f81a81ca566081e20dd1ff5f823cbe plus custom modifications.

Warning: This model is intended for mature audiences and may generate NSFW content. 不适合工作场景。 This repository contains a Qualcomm QNN/NPU optimized model converted from the One Obsession v24 Illustrious checkpoint, merged with DMD2 acceleration LoRA for mobile local image generation with Local Dream in NPU mode. - 运行软件：Local Dream - 运行模式：NPU 模式（Qualcomm Snapdragon） - 芯片适配：Qualcomm Snapdragon 8 Gen 3 及以上 - 转换工具：NPU Forge - Steps（步数）：8 ~ 12（推荐 10） - CFG Scale：1.0 - Scheduler（调度器）：LCM - Prompt Token Limit：220 tokens（提示词上限） - 光影强化：(chiaroscuro, high contrast:0.5), (sunbeam) - 变暗：(dark, chiaroscuro, high-contrast:1.7) - 半写实：((realistic, photorealistic, realistic shading)) - 还原动漫：(anime…

Model

onwx

Ryu Gyo

LLM を AMD Ryzen AI の NPU（XDNA2）だけで 動かすアプリです。行列計算から DeltaNet・アテンション・LM ヘッドまで、 1 トークンの計算をすべて NPU の上で行います（CPU は単語の表引きと次の単語の選択だけ）。NPU 向けのプログラムは OpenAI 互換 API サーバーとブラウザのチャット画面、タスクトレイのアプリ（モデルのダウンロード・切り替え・設定）を備えています。 AMD Ryzen AI の NPU で LLM を動かす代表的なアプリ FastFlowLM（FLM）1.0.8 と、 同じ測り方で比べました（Qwen3.5 9B、Ryzen AI Max+ 395、NPU は performance モード）。プロンプトは FLM の flm bench と同じ文章（長さ 128〜32K トークン）で、プロンプトを読み終わるまでの時間（TTFT）と、そのあと 32 トークンの生成を測っています。 点線は、設定の「CPU も使ってプロンプトの読み込みを速くする」をオンにした値です（既定はオフ）。 生成速度はどちらも MTP なし（温度 0 で 1 トークンずつ）の値です。FLM の Qwen3.5 には MTP がないので、条件はそろっています。 onwx の MTP（温度 0 のときの先読み）を使った値は、生成のグラフに破線で示しました。 - 512 トークンの回答をプロンプトから書き終えるまでの実効速度（512 ÷ 合計時間）は、128〜32K のどの長さでも onwx が上です（MTP なしでも）。…